See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221757566

Machine learning and pattern classification in identification of indigenous

retinal pathology

Article  in  Conference proceedings: ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine
and Biology Society. Conference · August 2011
DOI: 10.1109/IEMBS.2011.6091471 · Source: PubMed

CITATIONS READS

20 127

5 authors, including:

Herbert F Jelinek Anderson Rocha

Charles Sturt University University of Campinas
349 PUBLICATIONS   4,279 CITATIONS    160 PUBLICATIONS   2,050 CITATIONS   

SEE PROFILE SEE PROFILE

Tiago Carvalho Jacques Wainer

Federal Institute of São Paulo - Campinas, São Paulo, Brazil University of Campinas
37 PUBLICATIONS   248 CITATIONS    200 PUBLICATIONS   2,393 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Morphological changes in cardiac muscle associated with progression of Diabetes Mellitus View project

Automated Data-Driven Screening of Diabetic Retinopathy View project

All content following this page was uploaded by Herbert F Jelinek on 21 May 2014.

The user has requested enhancement of the downloaded file.

33rd Annual International Conference of the IEEE EMBS
Boston, Massachusetts USA, August 30 - September 3, 2011
Machine Learning and Pattern Classification in Identification of
Indigenous Retinal Pathology
Herbert F. Jelinek, Member, IEEE, Anderson Rocha, Member IEEE, Tiago Carvalho,
Siome Goldenstein, Senior Member, IEEE, Jacques Wainer
Abstract—Diabetic retinopathy (DR) is a complication of
diabetes, which if untreated leads to blindness. DR early
diagnosis and treatment improve outcomes. Automated
assessment of single lesions associated with DR has been
investigated for sometime. To improve on classification, especially
across different ethnic groups, we present an approach using
points-of-interest and visual dictionary that contains important
features required to identify retinal pathology. Variation in
images of the human retina with respect to differences in
pigmentation and presence of diverse lesions can be analyzed
without the necessity of preprocessing and utilizing different
training sets to account for ethnic differences for instance.
I. INTRODUCTION
I NDIGENOUS populations such as the Australian
Aborigine, all have an increased incidence of diabetes
compared to the Caucasian population resident in these
countries [1]. To optimize screening, detection and treatment,
mobile screening combined with automated classification of
disease can be used [2].
Automated assessment of pre-proliferative diabetic
retinopathy has been possible for some time using fluorescein-
labeled images [3]. Results for color fundus analysis
identifying microaneurysms, exudates and cotton-wool spots
as well as proliferative retinopathy have only been reported
more recently [4-6].
To optimize automated processing of color images one has
to consider intra-image variation such as light diffusion,
pathology, variation in fundus reflectivity and fundus
thickness and inter-image variation (being the result of using
different cameras, illumination, acquisition angle and retinal
pigmentation). Several methods are available including grey
world normalization, histogram equalization and histogram
specification [7]. Color normalization also increases the
discrimination in almost all of the color features [8].
Manuscript received March 26, 2011. This work was supported in part by
the Microsoft Research and Fapesp under the grants MSR-Fapesp
2008/54443-2 and Fapesp 2010/05647-4.
A. Rocha, T. Carvalho, S. Goldenstein, and J. Wainer are with the
Reasoning for Complex Data Lab., Institute of Computing, Univ. of Campinas
(Unicamp), Campinas, SP, Brazil. Contact (phone/fax: +55 19 35215854; e-
mail: anderson.rocha@ic.unicamp.br).
H.F. Jelinek is with the Centre for Research in Complex Systems and the
School of Community Health, Albury, NSW, Australia (phone: +61 2
60519219; fax: +61 2 60519219; e-mail: hjelinek@csu.edu.au).
978-1-4244-4122-8/11/$26.00 ©2011 IEEE 5951
This paper proposes a process that does not require
preprocessing but deals with image differences of the retinal
fundus directly. The approach constructs a visual dictionary to
represent important features that characterize the pathology of
interest and uses pattern recognition tools to classify the retinal
images into disease and non-disease.
II. METHODS
A. Diabetic Retinopathy Images
For creating the visual words and training the detectors
(two-class classifiers), 672 non-DR images, 261 images with
bright lesions and 246 images with red lesions from the
Ophthalmology Dept., Federal University of São Paulo, Brazil
were used. The images came from patients with different racial
background and were manually graded by specialists. The
images ranged from 640×640 to 1,581×1,113 pixels in
resolution. All of these images were used for creating the
dictionaries and for training the two-class classifiers (normal
vs. bright lesions and normal vs. red lesions). The aboriginal
test images were obtained from the Albury Eye Clinic using a
Topcon camera at 1,200×1,200-pixel resolution. Retinal
images contained either no pathology, red lesions, or bright
lesions.
B. Detection of Features
Every image in a collection is represented using a large
number of points of interest (PoI) [9] and a local descriptor
around each PoI using the Speeded-Up Robust Features
(SURF) approach is calculated [10]. This is then stored in an
indexing data structure. PoIs are robustly estimated as they
convey more information than other points in the image. A PoI
shows repeatability after several image transformations across
different scales.
The SURF algorithm has four major stages:
(1) Feature point detection: this stage uses an Hessian
detector approximation based on low-pass box filters (Haar
filters) and integral images [11] to speed up the operations.
(2) Feature point localization: The determinant of the
Hessian for both location and scale is used. Given a point
in an image I, the Hessian matrix in at
scale is defined as follows

(1)
where is the convolution of the Gaussian second
order derivative with the image in point .
The scale-spaces are implemented as image pyramids by
repeatedly smoothing the images with a Gaussian and
subsequently sub-sampling to achieve a higher level of the
pyramid. To localize the PoIs in the image across different
scales, the method performs non-maximum suppression in a
neighborhood. The maxima of the determinant of the
Hessian matrix are then interpolated in scale and image space.
(3) Orientation assignment: SURF calculates the Haar-
wavelet responses in and directions using a circular
neighborhood of radius around the interest point, where
is the scale at which the interest point was detected.
Fast filtering, in each scale , the method calculates wavelet
responses using integral images. The dominant orientation is
estimated by calculating the sum of all responses within a
sliding orientation window covering an angle of . The point
of interest gets the orientation from the longest vector.
(4) PoI characterization: SURF creates a square region
centered on the PoI, and oriented along the orientation selected
in Step (3). The region is split up regularly into smaller
square sub-regions. For each sub-region, the method computes
some simple features at regularly spaced sample points.
Basically, Stages (1) and (2) gives the points of interest (PoIs)
while Stage (3) assigns the orientation of each PoI and Stage
(4) performs the description of each PoI.
C. Visual Vocabulary
The creation of the dictionary is outlined in Fig. 1. SURF is
a good low-level representative feature detector with several
applications in computer vision. However, SURF-based
approaches are often designed to provide exact matching and
they do not translate directly into good results for image
classification in broad or even constrained domains. Therefore,
we use the concept of visual vocabularies [12] to capture the
high-distinctiveness of PoIs while using such discrimination
for image classification.
In the construction of a visual vocabulary, each region of
PoIs becomes a visual “word” of a “dictionary”. In the
following, we consider the problem of exudate detection for
the sake of explanation. The approach we discuss in this paper
is general enough to detect other DR-related anomalies as we
show in Section IV and V.
To solve the problem of detecting bright lesions in ocular-
fundus images, we select and create a database of training
examples comprising training positive images with exudates
and negative images considered normal by specialists. In this
training stage, we perform the localization of the interest
points in all available images using SURF. In this work, we do
5952
not perform any preprocessing on the images.
PoI Finding Example Image
PoI Description
(e.g., SIFT or SURF)
Vocabular y
Creat ion
(e.g., Clustering or with
Images with Anomaly
(e.g., Exudates)
Normal Images {20, 23, 1, 137, ...}
aid of Specialist)
Aa
D at abase of Visual Dictionar y
T raining Exam ples
Quantization
Test ing St age
2-Class ML Classifier
Input Example
to Classify PoI Finding
Description (e.g., SVM)
Aa
Test ing T r aining
Quantization Using the
Feat ure Vect or
? Feat ure Vect ors
Pre-Computed Dictionar y Classificat ion
Out come
Fig. 1.General pipeline of the proposed approach.
Each image in the training generates a series of points of
interest. After finding the PoIs, the dictionary or codebook was
created, which represents distinctive regions of the images
with bright lesions as well as images tagged as normal by
specialists. Our objective when creating a visual dictionary is
to learn, from a training set of examples, a model that selects
the more representative regions for our problem. The size of
our dictionary must be large enough to distinguish relevant
changes in the images and disregard irrelevant features.
To create the “dictionary”, we need to choose its size , or
number of representative words. During training, the
specialists select regions of interest in the analyzed images and
creating masks for candidate regions more likely to contain the
DR anomaly of interest. The points of interest are then
considered in these more likely DR regions. To create the
dictionary, we can then perform clustering such as k-means for
finding representative centers for the cloud of PoIs or simply
pick PoIs within the specialist marked regions as we indeed do
in this paper.
A good dictionary is the one that captures properties of the
DR anomaly of interest as well as properties of normal images.
Therefore, we create the dictionary using 50 “words”
representing the DR anomaly of interest (e.g., bright lesion)
and 50 “words” representing normal retinas (non-DR) [12].
Note that in the training stage we used common images and
not specific indigenous images.
D. Training and Classification
The visual dictionary was created from the training images,
using the fine selection of candidate regions in normal and
abnormal images, which the specialists marked. The selection
process is only performed in the training stage. Once the
dictionary was created, each of PoIs from the training set was
assigned to the closest visual word of the dictionary. This step
is known as quantization. At the end of the quantization
process, a set of feature vectors representing the histogram of
the selected visual words for each image are obtained. The
final classification procedure was performed using the Support
Vector Machine (SVM) algorithm. The classifier was trained
by adding the feature vectors calculated from the training
images containing positive (e.g., images containing pathology)
and negative (normal images) examples. The parameters for
the SVM were settled during training using SVM grid search.
To analyze multiple pathologies, we can create different
dictionaries (normal vs. bright lesions, normal vs. red lesions)
and train different two-class classifiers that can be combined
later using state-of-the-art machine learning classifier fusion.
III. RESULTS AND DISCUSSION
In this section, we present the results for our classification
techniques based on points of interest and visual words. Fig. 2
shows typical signatures for normal images vs. bright lesions.
Fig. 2.Typical signatures for normal patients vs. patients with
exudates. DR patients have higher frequencies for abnormal
words (positions 1-50) while the contrary should happen for
positions 51-100.
The plot depicts 100 “words” and their frequency in the
training set. For this research we chose to use 100 words as
this number was found to be most efficient and effective for
identification of lesions. Increasing the number of words
normally gives higher responses and also increases the
computational time. Thus positions 1-50 represent anomaly-
based words (e.g., exudates in this case) while positions 51-
100 represent words for normal regions (non-DR). For
positions 1-50, it is expected that the abnormal words
dominate the normal words while the contrary should happen
for positions 51-100.
The typical signatures for bright and red lesions are
calculated based on the training images. Given an unseen
aboriginal image for testing, the process consists of calculating
the points of interest in this image, and mapping such points to
the proper visual dictionary (normal and with anomalies).
Fig. 3 depicts two aboriginal images for bright lesions
detection/classification.
5953
Fig. 3 (top) shows an example of a good characterization of
an aboriginal image in which traces of bright lesions can
clearly be seen. This is due to the dominance of the tested
signatures with respected to the typical signature of a DR
patient with bright lesions (higher frequencies in positions 1-
50).
Fig. 3. Two examples of visual words characterization for DR
aborigine bright lesions detection. Best characterization results
(top) against worse characterization results (bottom).
On the other hand, Fig. 3 (bottom) shows a bad
characterization in which it is not possible to differentiate
between a DR candidate for bright lesion since the signature
found in the tested image does not dominate the typical
signature for DR or normal patients. Fig. 4 presents similar
results for red lesions. Fig. 4 (top) indicates that the tested
image is a DR candidate for red lesion since its signature for
positions 1-50 dominates the typical red lesions signature
while for positions 51-100 the typical normal signature
dominates. Fig. 4 (bottom) depicts an example in which the
red lesions characterization is not as good and no clear
conclusion can be drawn.
This kind of analysis is very important to visualize what is
happening in the feature level characterization of the analyzed
images, but difficult to automate. Therefore, use is made of a
typical machine learning classifier to learn such behaviors and
 test unknown images. For this we selected an operational point
in which false negatives, for instance, were penalized.
Fig. 4. Two examples of visual words characterization for DR
aborigine red lesions detection. Best characterization results
(top) against worse characterization results (bottom).
The classification then achieved promising results, correctly
classifying about 80% of the aboriginal images from a training
set of non-aboriginal images. The two-class SVM classifier
with a radial basis kernel was used with class weighting during
training to balance differences in the number of examples for
normal, bright, and red-lesion images, and at the ROC
operational point of 90% sensitivity and 80% specificity.
This is an important result as it indicates not only that the
method works on indigenous images but also that cross
training has not reduced the accuracy of the procedure. Having
a common approach which is robust to cross-training can be an
important milestone in DR pathology detection research.
IV. CONCLUSION
The use of the visual dictionary is a robust method to learn and
represent important features of a given anomaly even in the
presence of noise and differences in color background of the
retinal fundus in different populations.
Ocular-fundus images were classified as normal or DR
5954
View publication stats
candidates using a cross-training methodology that was robust
against differences in retinal fundus color and therefore easy to
implement worldwide. In addition, the unified proposed
approach allows us to develop different detectors under the
same simple underlying characterization procedure allowing
more than one lesion to be identified at a time therefore
yielding better differentiation in terms of DR disease
progression. As a future direction, we aim at investigating
machine learning feature and classifier fusion techniques in
order to combine different anomaly detectors toward more
discriminative DR vs. Non-DR classifiers.
ACKNOWLEDGMENT
We thank the Federal University of São Paulo medical team
for collecting and grading the ocular-fundus images. In
particular, we express our gratitude to Dr. Eduardo Dib, from
Federal Univ. of São Paulo, for taking the lead in grading the
training images as well as to Dr. Alan Luckie, from the Albury
Eye Clinic, for providing the aboriginal images and initial
classification of the testing retinal images.
REFERENCES
[1] P. Mitchell, S. Foran, T. Y. Wong, B. Chua, I. patel, and E. Ojaimi,
Guidelines for the Management of Diabetic Retinopathy. Canberra:
NHMRC, 2008.
[2] H. F. Jelinek, D. Cornforth, M. Cree, J. Cesar, R.M., J. J. G. Leandro, J.
V. B. Soares, and P. Mitchell, "Automated characterisation of diabetic
retinopathy using mathematical morphology: a pilot study for
community health," in NSW Primary Health Care Research and
Evaluation Conference, Sydney, 2003, pp. 48.
[3] M. J. Cree, J. A. Olson, K. McHardy, P. Sharp, and J. Forrester, "A fully
automated comparative microaneurysm digital detection system.," Eye,
vol. 11, pp. 622-628, 1997.
[4] A. L. Karperien, H. F. Jelinek, J. J. G. Leandro, J. V. B. Soares, J. Cesar,
R.M., and A. Luckie, "Automated detection of proliferative retinopathy
in clinical practice," Clinical Ophthalmology, vol. 2, pp. 109-122, 2008.
[5] H. Wang, W. Hsu, K. G. Goh, and M. L. Lee, "An effective approach to
detect lesions in colour retinal images," in IEEE Int. Conf. in Computer
Vision and Pattern Recognition, 2000, pp. 181-187.
[6] L. Streeter and M. J. Cree, "Microaneurysm detection in colour fundus
images," in Image and Vision Computing, 2003, pp. 280-284.
[7] K. A. Goatman, M. J. Cree, J. A. Olson, P. F. Sharp, and J. V. Forrester,
"Automated measurement of microaneurysm turnover," Investigative
Ophthalmology and Visual Science, vol. 44, pp. 5335-5341, 2003.
[8] M. J. Cree, E. Gamble, and D. J. Cornforth, "Colour normalisation to
reduce inter-patient and intra-patient variability in microaneurysm
detection in colour retinal images," in Workshop on Digital Image
Computing, Brisbane, Australia, 2005, pp. 163-169.
[9] E. Valle, M. Cord, and S. Philipp-Foliguet, "High-dimensional descriptor
indexing for large multimedia databases," in ACM Intl. Conf. on
Information and Knowledge Management, 2008, pp. 739–748.
[10] H. Bay, T. Tuytelaars, and L. V. Gool, "SURF: Speeded up robust
features," in European Conf. on Computer Vision, 2006, pp. 1-14.
[11] P. Viola and M. Jones, "Robust real-time face detection," Intl. Journa of
Computer Vision, vol. 52, pp. 137–154, 2004.
[12] A. Rocha, T. Carvalho, S. Goldenstein, and J. Wainer. "Points of
Interest and Visual Dictionary for Retina Pathology Detection,"
Technical Report IC-11-07, Institute of Computing, Univ. of Campinas,
Campinas, Brazil, 2011.